Kimi K3

Moonshot AI launched Kimi K3 on 16 July 2026 without a complete official benchmark table. The specifications are vendor-stated: a mixture of experts of roughly 2.8 trillion parameters, a one-million-token context, native multimodality, and Kimi Delta Attention, a hybrid linear attention credited with decoding 6.3 times faster at one million tokens. SFEIR read the launch as engineers, treating those numbers as claims and the community arena scores as indications.

Mozilla's July 2026 report places the model at 57 on the Artificial Analysis Intelligence Index, fourth overall, against 61 for the best closed model, Claude Opus 5. On the Epoch Capabilities Index the gap is six points, which Mozilla calls about one release cycle, with overlapping confidence intervals.

Z.ai measured itself against the same model in August 2026. Its GLM-5.3 announcement claims the title of most capable open-weights model for coding, yet Kimi K3 leads on Terminal Bench 2.1, DeepSWE v1.1, SWE-Marathon v1.1 and Toolathlon Verified. The head-to-head reads 3 to 3 with one tie, so the claim does not follow from the table.

Price is where SFEIR located the pressure: around $3 per million input tokens, $15 output, $0.30 cached, figures drawn from early reviews that still need verification. A frontier model at open weights and that price pulls the price-performance curve down. Kimi K3 replaces Kimi K2.5 (fermé aux nouveaux, extinction au 31 août 2026) and adds a reversibility column to multi-model routing, at the cost of running 2.8 trillion parameters yourself.